If the LLM Can't Find You, You Don't Exist Discoverability for MCP Apps - Vincent McLeese

In this talk, Ghost Team CTO Vincent McLeese shares the three foundations every app builder needs to get their ChatGPT and MCP apps discovered, chosen, and used - before competitors take their spot. We analyzed 148 live ChatGPT apps and ran over 15,000 simulations to crack open the black box of MCP app discoverability. 📌 Get your MCP App & ChatGPT App discovered: https://www.appdiscoverability.com 📌 MCP App Intelligence: https://track.appdiscoverability.com 📌 Join the #1 ChatGPT App Builder Community: / discord Here's what we cover: → The Discoverability Problem: 14% of the time, ChatGPT apps fail to invoke on obvious, direct prompts. That's 1 in 7 — imagine if your website just didn't load that often. → How MCP App Discovery Works Today: The current state across OpenAI, Claude, and Microsoft Copilot — from manual connect-and-mention to fully organic discovery. → Live LLM Monitoring: Why MCP server logs aren't enough and what you're missing — model downgrades, geography, personalization, and invocation failures you can't see. → Model Degradation Effects: How the exact same prompt works on one model but fails on another, and what happens when free users get silently downgraded. → Mentioning ≠ Invocation: Even when users directly mention your app, the model may ignore your tools in favor of its own training data. Why tool descriptions matter more than you think. → Conversational Testing: Golden prompt sets are useless. Real users prompt messily — and we found a 20-point invocation difference between polished and non-technical prompts across 5,000 tests. → Continuous Optimization of Tool Descriptions: Stanford research found almost 100% of tools have quality defects. For Statista, a single word change improved tool success rate by nearly 6%. → The SEO Parallel: In 2005 we one-shotted metadata for websites. That's exactly what we're doing now with tool descriptions — and where it's headed is autonomous, agentic-driven optimization. → Competitive Category Analysis: How six travel apps performed with increasingly vague prompts, and why clear "use this when" language in tool descriptions wins. → What Ghost Team Is Building: App Discoverability platform combining LLM monitoring, conversational testing, and continuous tool metadata optimization — now in beta. MCP app discoverability is the new SEO. The builders who start monitoring and optimizing now will own the organic discovery layer when it goes fully live. Don't wait. 📌 Ghost Team: https://www.ghostteam.ai 📫 Newsletter: https://elliotgarreffa.beehiiv.com/ Follow for More AI Content: ➡️ LinkedIn:   / elliotgarreffa   ➡️ X: https://x.com/elliot_garreffa ➡️ TikTok: /elliotagarreffa ➡️ Instagram: /elliotagarreffa Timestamps 00:00 – Intro: Who is Ghost Team and what is App Discoverability 01:30 – The setup: You've submitted your app — now what? 03:00 – The black box: 148 apps analyzed, 15,000+ simulations 04:30 – Quick MCP app recap: Over a billion users, 600+ apps tracked 06:30 – How discovery works today: OpenAI, Claude, and Copilot compared 09:00 – The discoverability problem: 14% failure rate on obvious prompts 10:30 – Foundation 1: Live LLM monitoring 13:00 – Model degradation and silent downgrades 14:30 – Why mentioning your app doesn't guarantee invocation 16:00 – Foundation 2: Conversational testing 18:00 – Golden prompts vs. real user behavior: The 20-point gap 20:00 – Foundation 3: Continuous optimization 22:00 – Stanford research: Almost 100% of tools have quality defects 23:30 – The Statista case: One word = 6% improvement 25:00 – Competitive category analysis: Travel app tool descriptions compared 27:00 – The SEO parallel and the future of agentic optimization 29:00 – Ghost Team + App Discoverability platform beta 30:00 – Q&A: The history of organic discovery on ChatGPT